Studies on multi-terminal charge trap memristive synaptic devices for energy efficient neuromorphic system

Abstract

Neuromorphic computing is the next generation of artificial intelligence that mimics brain functions, providing dramatic energy efficiency for processing and storing data. Software-based research has made many advances in the field of existing neuromorphic research, but hardware-based research is somewhat insignificant. Using a charge trapping-based system, I implement (i) a two-terminal-based synaptic device that follows Hebbian learning rules, (ii) a three-terminal-based synaptic device that follows BCM learning rules, and (iii) a synaptic device that shows the retinal nerve simulation characteristics of mantis shrimp with the best visual system. First, the Nb$_2$O$_5$-based charge trap memristor (CTM) exhibits self-rectifying, low-current, electroforming-free, and analog switching characteristics. I fabricated a CTM with an array size of 32×32 array comprised of the Pt/HfO$_2$/Nb$_2$O$_5$/HfO$_2$/Ti structure. They exhibited all the desirable features (self-rectifying, low-current, electroforming-free, and analog switching characteristics) with all devices working. Furthermore, all memory cells in the array provided four bits (16 levels) stably between 0.45 nS and 7.45 nS at 3 V of reading voltage, allowing me to build a neuromorphic hardware testing platform. Second, an oxide-based memtransistor was manufactured by combining the material technology of the two-terminal Nb$_2$O$_5$-based CTM and 3-terminal ITZO (In-Sn-Zn-O)-based thin film transistor (TFT) structure. Memtransistor is a suitable candidate for synaptic devices that can modulate BCM learning rules. I used two methods to create a membrane transistor structure in an oxide-based device to demonstrate a three-terminal ITZO-based FET using a charge-trapping memristive system imitating the triplet STDP. A memristor in the direction of horizontals was implemented to develop a leaky IGZO-based FET through active carrier doping. In addition, another memristor in the direction of vertical was implemented through the post-deposition of the charge trap layer (Ta$_2$O5/Nb$_2$O$_5$-X/Al$_2$O$_3$-X) deposited on an active layer, where used in 2-terminal charge trap memristor (CTM). It was used in Triplet-STDP as a suitable candidate for synaptic devices that can mimic BCM learning rules by being able to adjust the lateral memristor in the vertical direction. Third, A three-terminal-based photonic synapse imitating mantis shrimps was produced using defect-engineered Ag$_2$S chiral quantum dots (QDs). Mantis shrimps see the world through multiple photoreceptor channels. The optical information includes light's wavelength and circular polarization, making them have the best eye in nature. A chiroptical synapse (ChiropS) shows different polarized light absorption behaviors at different wavelengths, suggesting their applicability for multi-channel optical sensing, recognizing approximately 5% difference at 405 nm wavelength and 6% difference at 532 nm wavelength. We integrated a photonic synaptic device that readily detects and distinguishes the polarized light. Moreover, its combined short- and long-term plasticity after illumination allows it to perform the pre-processing function for image-noise filtering, reducing energy consumption by 21% on the Gaussian noise-added handwritten MNIST dataset. These studies will add a new dimension to next-generation processing devices with high integration and excellent energy efficiency.